Computer-aided system for diagnosis of neurological disorders

Abstract

A non-invasive computer-aided system for diagnosis of neurological disorders uses as input neuroimaging data of a subject brain, parcellates the subject brain into a plurality of brain regions, determines markers of functional connectivity between pairs of brain regions, and uses a machine learning classifier to classify the subject brain with respect to a neurological disorder based on the determined markers of functional connectivity.

Description

COMPUTER-AIDED SYSTEM FOR DIAGNOSIS OF NEUROLOGICALDISORDERS

[0001] This application claims the benefit of United States provisional patent application serial no. 63580458, filed September 5, 2023, for COMPUTED-AIDED SYSTEM FOR DIAGNOSIS OF NEUROLOGICAL DISORDERS, incorporated herein by reference.FIELD OF THE INVENTION

[0002] A non-invasive computer-aided system for diagnosis of neurological disorders uses as input neuroimaging data of a subject brain, parcellates the subject brain into a plurality of brain regions, determines markers of functional connectivity between pairs of brain regions, and uses a machine learning classifier to classify the subject brain with respect to a neurological disorder based on the determined markers of functional connectivity.BACKGROUND OF THE INVENTION

[0003] Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by three primary characteristics: functioning difficulties with social interaction, communication barriers, and behavioral restrictions and repetitive patterns. Despite the lack of a comprehensive understanding of ASD causes, numerous hypotheses and theories have been proposed concerning the etiology of its underlying mechanism. These hypotheses and theories suggest that genes and environmental factors play a significant role in determining ASD severity. Anatomical abnormalities of the brain, functioning of the brain during rest or while performing different tasks, or abnormal connectivity of the brain’s white matter are hypothesized to be responsible for ASD symptoms. Several magnetic resonance imaging (MRI)-based imaging methods have been utilized to study a variety of abnormalities correlated with ASD, including: (i) structural magnetic resonance imaging (sMRI) for anatomical anomalies; (ii) functional magnetic resonance imaging (fMRI), either at rest or while performing a task, for abnormalities in brain activity; and (iii) diffusion tensor imaging (DTI) for abnormalities in connectivity. Brain connectivity refers to the strength of interactions, whether director or indirect, between different brain areas which locally process information.

[0004] In resting state fMRI (rs-fMRI), brain activity is captured, in terms of blood-oxygen- level-dependent (BOLD) signals, while the subject is at rest. Since connectivity patterns reveal the resting state, they have proven to be beneficial in diagnosing mental disorders, such as schizophrenia and Alzheimer’s disease. There is evidence that the functional connectivity between major brain networks, as well as the functional connectivity within individual networks, is altered in ASD, as demonstrated by numerous studies. Certainstudies have found that people with ASDs have reduced functional connectivity, which supports the hypothesis that autism is characterized by a deficit in functional connectivity, or ’underconnectivity’. In accordance with the underconnectivity theory, cognitive disorders manifest themselves in reduced synchronized brain activity during integrative processing tasks, such as the synthesis of a sentence from a set of words. There is evidence in previous literature for the hypothesis of underconnectivity, suggesting lower functional connectivity in the superior parietal and visuospatial regions of ASD when compared to typical development (TD), reduced connectivity between the temporal and frontal cortex without global abnormalities, and abnormalities in functional networks, which were more evident in networks related to social information processing. However, other studies have indicated that brain networks associated with ASDs exhibit both under-connectivity and overconnectivity - increased brain connectivity in some areas when compared to healthy controls. In such studies, functional connectivity patterns were analyzed, specifically interhemispheric connectivity analysis, and it was found that subjects with ASD displayed both under- and over-connectivity between different brain regions. Another study identified hyperconnectivity in severely socially challenged autistic children. Accordingly, ASD remains a complex neurological disorder not readily diagnosed based on rs-fMRI data.

[0005] The term “curse of dimensionality” has been used to describe the problem of exponential complexity resulting from the addition of new dimensions to a dataset (also referred to as addition of features to feature space), commonly defined as the p » n problem. MRI imaging research and medical data are commonly affected by this phenomenon, resulting in over-fitting. There are several ways to reduce the number of features, such as principal component analysis, linear discriminant analysis, or autoencoders. Due to the fact that these methods do not preserve the semantics of the original feature space, it is typically difficult to identify what clinical findings underlie such classification results. This, in turn, makes them less useful for actual use by physicians to provide them with more information or to better understand the pathological abnormalities underlying each autistic brain and, thus, are less attractive and less practical.

[0006] Due to autism’s heterogeneous nature, brain connectivity can vary widely among individuals with the disorder, making classification of the condition difficult. The design of a less-sensitive functional connectivity feature that is less affected by age, sex, and designs of the resting-state scan and studying its correlation with autism is an active research area, emerging into two popular brain connectivity representations: the conventional, more popular, static functional connectivity (FC), and the newer dynamic functional connectivity (dFC) representations. Those functional connectivity metrics are believed to capture different internal states of the brain while at rest. The static FC is a matrix obtained by calculating the Pearson cross-correlation coefficient of the BOLD signals across pairs of pre-defined brainareas. Following statistical analysis and depending on whether one examines local or global networks, there is evidence of both under- and over-connectivity in autism in the majority of the literature. As most of the literature is directed to temporally stationary functional networks in resting state, dynamic or time-varying changes to functional connectivity that occur during brain scanning are not sufficiently considered by static FC. Newer studies suggest periodically changing spatial patterns of functional networks. To capture those dynamic connectivity patterns, multiple computational strategies were used to find dFC that consider such temporal fluctuations of functional connectivity. dFC analyses allow identification of not only common brain states but also transitions between them. The most commonly used approach for dFC computation is sliding window techniques, while other approaches include clustering methods, dynamic connectivity regression, time-frequency analysis, wavelet transforms, dynamic connectivity detection, and time series models.

[0007] A significant limitation of previous works in the field is that each only gives a partial view of the essential components and foundational details of an efficient computer-aided diagnosis (CAD) system. As the characteristics of autism vary from individual to individual in terms of symptoms and severity, a more personalized approach to predicting and analyzing the behavior and functional capabilities of each autistic subject has become increasingly necessary.SUMMARY

[0008] The instant subject matter relates to a non-invasive computer-aided system for diagnosis of neurological disorders uses as input neuroimaging data of a subject brain, parcellates the subject brain into a plurality of brain regions, determines markers of functional connectivity between pairs of brain regions, and uses a machine learning classifier to classify the subject brain with respect to a neurological disorder based on the determined markers of functional connectivity. Functional connectivity between a pair of brain regions is established by the Pearson correlation coefficient between the average time courses of the two brain regions. The average time course for each region is calculated as the average of BOLD of all corresponding voxels over the time in this region.

[0009] In some embodiments, the present invention includes a system for ASD diagnosis based on the analysis of brain rs-fMRI data and incorporates the advantages of new dynamic feature representation and machine learning prospects. The CAD system provides an accurate state-of-the-art diagnosis of ASD based on a public dataset, and can also be used for identifying the brain regions contributing to such diagnosis. While a human physician may make the ultimate diagnosis, the CAD system can provide an informed prediction of the diagnosis to aid the physician’s diagnosis. Having a better understanding of the brain abnormalities associated with autism is another aspect of this system. A furtheradvantage of the system is that it easily scales up: more subjects can be preprocessed and their features calculated independently, and an additional mode of imaging over the same subject, such as structural MRI, can be incorporated into the feature selection stages.

[0010] It will be appreciated that the various systems and methods described in this summary section, as well as elsewhere in this application, can be expressed as a large number of different combinations and subcombinations. All such useful, novel, and inventive combinations and subcombinations are contemplated herein, it being recognized that the explicit expression of each of these combinations is unnecessary.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] A better understanding of the present invention will be had upon reference to the following description in conjunction with the accompanying drawings.

[0012] FIG. 1 depicts a schematic illustration of the steps of the CAD system of present invention.

[0013] FIG. 2 is an illustration of the brain extraction step from fMRI image data depicting an original fMRI image (left panel) and the extracted brain (right panel).

[0014] FIG. 3 is an illustration of the brain parcellation step from fMRI image data depicting the MNI152 standard space (left panel) and the parcellated extracted brain (right panel - neuroimaging data, post-parcellation, typically distinguishes different brain regions by color, depicted here in greyscale).

[0015] FIG. 4 is a visual diagram of the calculations of the static functional connectivity FC representation, extracted from two different brain areas.

[0016] FIG. 5 is a visual diagram of the calculations of the dynamic functional connectivity dFC representation, extracted from two different brain areas.

[0017] FIG. 6 is an illustration of estimated synchronization among all pairs of brain regions (left panel) and the identified brain regions linked to ASD (right panel). Each row and column is indicative of a brain region, wherein a brighter color at the intersection of a row and column is indicative of greater synchronization between the regions associated with that row and that column.DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] The details of one or more embodiments of the presently-disclosed subject matter are set forth in this document. Modifications to embodiments described in this document, and other embodiments, will be evident to those of ordinary skill in the art after a study of the information provided in this document. The information provided in this document, and particularly the specific details of the described exemplary embodiments, is provided primarily for clearness of understanding and no unnecessary limitations are to be understoodtherefrom. In case of conflict, the specification of this document, including definitions, will control.

[0019] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the presently-disclosed subject matter belongs. Although any methods, devices, and materials similar or equivalent to those described herein can be used in the practice or testing of the presently-disclosed subject matter, representative methods, devices, and materials are now described.

[0020] Following long-standing patent law convention, the terms “a”, “an”, and “the” refer to “one or more” when used in this application, including the claims. Thus, for example, reference to “a cell” includes a plurality of such cells, and so forth.

[0021] Unless otherwise indicated, all numbers expressing quantities of ingredients, properties such as reaction conditions, and so forth used in the specification and claims are to be understood as being modified in all instances by the term “about.” Accordingly, unless indicated to the contrary, the numerical parameters set forth in this specification and claims are approximations that can vary depending upon the desired properties sought to be obtained by the presently-disclosed subject matter.

[0022] As used herein, the term “about,” when referring to a value or to an amount is meant to encompass variations of ±10% of the most precise digit in the value or amount (e.g., “about 1” refers to 0.9 to 1.1, “about 1.1” refers to 1.09 to 1.11, etc.).

[0023] As used herein, ranges can be expressed as from “about” one particular value, and / or to “about” another particular value. It is also understood that there are a number of values disclosed herein, and that each value is also herein disclosed as “about” that particular value in addition to the value itself. For example, if the value “10” is disclosed, then “about 10” is also disclosed. It is also understood that each unit between two particular units are also disclosed. For example, if 10 and 15 are disclosed, then 11, 12, 13, and 14 are also disclosed.

[0024] The present invention relates to a computer-aided system for diagnosis of neurological disorders. While the present invention is discussed in terms of a computer- aided system for diagnosis of ASD, it should be understood that the methodology described herein may be used with other neurological disorders, such as, for example, attention-deficit I hyperactivity disorder (ADHD), schizophrenia, or others.

[0025] The present invention provides for creation of a treatment plan that is tailored specifically to the needs of each autistic individual. The system provides not only an accurate diagnosis but also identifies the brain regions contributing to the diagnosis.

[0026] In this work, the Autism Brain Imaging Data Exchange I (ABIDE I) and ABIDE II datasets are used, which are publicly available repositories with rs-fMRI and sMRI imagingmodalities. The ABIDE I database includes rs-fMRI data from 884 subjects, all of which were used in this project. Table 1 provides a summary of the 884 subjects’ demographics: age in years, label, full-scale IQ, and gender.Table 1: Demographics summary of the 884 subjects with rs-fMRI data, showing sex counts (M: male, F: female), age at scan in years, and full-scale IQ (FIQ)ASD Group (n = 408) TD Group (n = 476)M = 358, F = 50 M = 388, F = 88AGE FIQ AGE FIQ count 408 379 476 442 mean 17.69 106.19 16.79 111.28 std 8.93 17.01 7.35 12.48 min 7 41 6.47 73 max 64 148 56.2 146This project also involved use of a cohort of individuals from the ABIDE II dataset, 344 with ASD and 374 TD, for a total of 718 subjects. The subjects with ASD include 288 males and 56 females with a mean age of 13.67 and a standard deviation of 9.30. The subjects with TD include 267 males and 107 females with a mean age of 13.71 and standard deviation of 8.73.

[0027] FIG. 1 illustrates the framework of the CAD system, which includes the steps of (10) receiving as input one or more neuroimaging data of a subject brain, such as, for example, fMRI volumes and, optionally, sMRI volumes, (12) preprocessing of the input neuroimaging data, (14) extraction and parcellation of the subject brain into a plurality of brain regions, (16) generating local feature representations, wherein each feature representation is a determination of the functional connectivity between neuroimaging data of pairs of brain regions, (18) selecting feature representations characteristic of a neurological disorder, and (20) classifying, using a machine learning classifier, the subject brain with respect to the neurological disorder based on the feature representations characteristic of the neurological disorder. In some embodiments, the classifying step includes classifying the subject brain as indicative of the neurological disorder or typical development. In further embodiments, the classifying step includes classifying the subject brain as indicative of the neurological disorder at a specified severity level or typical development. In a particular embodiment, where the neurological disorder is ADS, the subject brain is classified as or diagnosed as indicative of mild ASD, moderate ASD, severe ASD, or typical development. Each step and the diagnostic accuracy of the classifications and resulting final diagnosis are described interms of a first embodiment, wherein various parameters are evaluated for the steps and the preferred parameters identified, and a second embodiment, wherein the CAD system is operated using the preferred parameters. The first embodiment provides a roadmap for developing a CAD system for diagnosis of a neurological disorder, with ASD used as the exemplary disorder, while the second embodiment provides a CAD system for diagnosis of ASD.

[0028] In a first embodiment, the preprocessing step includes preprocessing of functional MRI data according to the standard configurable pipeline for the analysis of connectomes (C-PAC). The preprocessing pipeline includes performing slice timing correction, motion realignment, skull stripping, and intensity normalization in this order. Twenty-four motion parameters were used as regressors to overcome the subject’s motion-related confounding variables. Additionally, five more from white matter and CSF mean levels were regressed to eliminate their effects. Then, one of four different preprocessing strategies is applied for filtration and signal correction. The strategies differ by whether global signal correction (i.e. , inclusion of global mean signal in nuisance regression) is made and whether filtration (i.e., band-pass filtering after global signal correction) is performed. Those four strategies are denoted as: filt_global (i.e., with inclusion, with filtration), filt_noglobal (i.e., without inclusion, with filtration), nofilt_global (i.e., with inclusion, without filtration), and filt_noglobal (i.e., without inclusion, without filtration). Lastly, structural MRI data, after skull stripping, segmentation, normalization, and registration to the standard MN I- 152 space, is used for registration of the corresponding fMRI volumes to the same space after those preprocessing steps. Each of the four preprocessing strategies generated different outputs, which were compared to determine to most effective strategy.

[0029] In a second embodiment, the preprocessing step includes preprocessing of functional MRI data using the existing FasterSurfer and fMRI Prep processing tools. This process includes the steps of: aligning fMRI with sMRI to enhance the localization of functional activations by augmenting fMRI resolution, employing head-motion estimation to the Strongen data quality by detecting and rectifying movement-related artifacts and distortions, conducting slice time correction to align slices from different time points and mitigate temporal discrepancies, implementing susceptibility distortion correction to rectify fMRI distortions stemming from magnetic susceptibility variations, and performing preprocessing of BOLD signals in native space to enhance BOLD signal data within its original acquisition space.

[0030] In a first embodiment, the extraction and parcellation steps include identifying different brain regions on the registered data using a standard atlas parcellation. Extraction was performed using conventional techniques. A standard atlas is used to localize the fMRI features (time-series BOLD signals) to different brain regions, as defined by each atlas. Insome embodiments, the disclosed CAD system uses two different atlases, such as, for example, the automated anatomical labeling (AAL) atlas, defining 116 brain regions and the Talaraich and Tournoux (TT) atlas, defining 97 different regions using Brodmann area labeling. In this work, all brain regions defined by each of the two used brain atlases are analyzed. For each atlas, mean BOLD signals for each parcellated brain region are extracted. In other embodiments, other brain region atlases or additional brain region atlases may be used.

[0031] In the second embodiment, extraction of the brain is accomplished by leveraging spatial information derived from sMRI, utilizing a probabilistic brain atlas, and considering the visual characteristics of brain tissues. The information obtained from these sources is then input into a Bayesian classifier, which aids in distinguishing between brain tissues and the skull, as illustrated in FIG. 2. The MNI152 standard space was used to segment the extracted brain into anatomically similar regions and establish correlations with ASD, enabling discrimination between ASD and TD brains. 3DC-based affine registration was used to align the extracted brain with the MNI152 standard space as illustrated in FIG. 3. The AAL atlas is then used to localize the fMRI features (time-series BOLD signals) to different brain regions.

[0032] After preprocessing, extraction and parcellation, an average time course for each brain region was calculated as the average of BOLD signals of all voxels corresponding to that region over time. In the first embodiment, eight datasets of average BOLD signals have been produced, one for each combination of the four preprocessing strategies and two atlases. In the second embodiment, a single dataset of average BOLD signals has been produced according to the described preprocessing strategy, MNI152 standard space, and AAL atlas, although other atlases, standard spaces, or preprocessing strategies may may be used in other embodiments. The next step is to convert the average BOLD signal to a more meaningful representation that captures functional connectivity and is less dependent on the subjects’ differences.

[0033] In the first embodiment, for the purpose of studying the coherence between different areas of the brain, two different feature representations were used to examine functional connectivity. First, the most famous and commonly used static FC matrix, where the correlation between the full duration of the activation courses is used as a measure of functional connectivity. The reason behind this selection is that it captures the intrinsic functional network of the brain.

[0034] The static functional connectivity matrix FC is constructed as the Pearson correlation coefficient (p) between each pair of the (A = 116 | 97 ) areas I brain regions in the atlas. The effective output feature size is (A x (A - 1) 2), because of symmetry, resulting in a total of 884 x 6670 for AAL, or 884 x 4656 for TT, feature matrices (i.e., 6670 for AAL or 4656 for TTfor each of 884 subjects in the ABIDE I dataset). FIG. 4 illustrates the pipeline of this adopted feature representation, wherein BOLD signals of representative brain regions i and j are correlated to generate a fMRI connectivity matrix.

[0035] The second feature representation introduced in this work is an enhanced version of dynamic functional connectivity, dFC, where temporal dynamics are considered in the correlation calculation. The calculation starts by multiplying a short Gaussian sliding window with width wwith each time signal, to calculate pair correlation, with an overlapping step size s. A Gaussian window of size w = 21 TR, o = 3TR, and a step size of s = 1 is used, where TR is the fMRI repetition time, typically 1500-2000 ms depending on the scanning site. Hence, for each pair of brain regions of length L, M correlations are calculated, where M = \\(L - w) / s||. Following the hypotheses of under-connectivity and over-connectivity differences between the functional activation of an autistic and typically developed brain, a quantification using both the percentage of strong correlations nsrand the percentage of weak / no correlation nw / <are used to represent over- and under-connectivity between each brain area, where0.25.This creates two metrics for each pair of regions identifying the existence of under- and overconnectivities, yielding double size dFC feature matrices in comparison to the FC representation. FIG. 5 illustrates the pipeline of the second feature representation in this system. At this step, sixteen different datasets have been generated in the first embodiment, one for each combination of preprocessing strategy, atlas, and feature representation (i.e. , 4 strategies x 2 atlases x 2 feature representations (static FC or dFC)), each being a model.

[0036] In the second embodiment, the step of generating a plurality of local feature representations is enacted using dFC to determination of the functional connectivity between neuroimaging data of pairs of brain regions as described above. As noted above, the effective output feature size is (A x (A - 1) 2), resulting in a total of 6670 for AAL feature matrices for each of the 718 subjects in the ABIDE II dataset. This approach aims to detect any potential correlations with ASD across a wide range of brain regions. Utilizing normalized cross as an image radiomic, the synchronization between each pair of interconnected brain regions may be represented graphically, as illustrated in FIG. 6, left panel.

[0037] A feature reduction technique is employed to address the curse of dimensionality problem and identify the feature representations characteristic of and diagnostically relevant to the specified neurological disorder. In the first embodiment, several feature reduction techniques were explored. Since the employed feature reduction shall preserve the semantics of original features, many popular techniques such as principal componentanalysis (PCA), factor analysis (FA), or linear discriminant analysis (LDA) may not be suitable, as they create a newly transformed feature space from which it is difficult to extrapolate the meaning of any following clinical classification, or underlying functional abnormalities of autistic individuals. Two cascaded feature selection stages are used to select the most significant or discriminant subset of features: a simple univariate selector followed by a more sophisticated recursive feature elimination (RFE) with cross-validation (RFE-CV).

[0038] The univariate selector is a fast method that aims to initially reduce the feature space to make the second stage computationally efficient. It computes the ANOVA F-value score for each feature to select the top proportion accordingly. The F-value is the ratio of the between group variation and the within group variation, wherein a large F-value means the between-group variation is larger than the within-group variation, which can be interpreted to mean there is a statistically significant difference in the means of each group. Following this, a RFE eliminates less significant (i.e. , weaker) features until a specified number of features is reached by fitting a kernel classifier and removing the less significant features. By eliminating dependencies and collinearities, RFE attempts to achieve a better understanding of the model. Cross-validation aims to score the significance (i.e., strength) of features on the test subset, other than the training folds data used in training the kernel, to mitigate overfitting. The step of the eliminated features, scoring metric, kernel type, and the number of folds are all parameters of choice. Here, a step size of 2 features (for faster elimination of less significant features, yet small enough for a good performance), k = 5 folds (common 5- fold cross-validation), and balanced accuracy were chosen. Four different classifiers were tested as a feature selection kernel, namely linear support vector machine (LSVM), logistic regression (LR), random forest (RF), and light gradient boost machine (LGBM), to investigate different feature-relationships. The two-stage feature selection is applied to each dataset / feature representation to find the best n features that provided the best crossvalidated score.

[0039] In the second embodiment, RFE was employed in tandem with 5 k-fold cross validation to identify the brain regions associated with ASD. FIG. 6, right panel, illustrates the brain regions that exhibit correlation with ASD.

[0040] In the first embodiment, performance of the sixteen models was next evaluated on a set of machine learning classifiers to select which model to use for further processing in order to learn how to classify autistic brains. A system of different machine learning classifiers was established from the set of n features selected for each of the sixteen models. Six different classifier types were evaluated and their hyperparameters optimized in order to establish the most accurate parameter classifier model. Those classifiers, including linear and non-linear ones, are: (i) Isvm; (ii) Ir; (iii) rf; (iv) Igbm; (v) neural networks (nn); and(vi) radial-basis function SVM. The different types would test different relations between the two feature classes, with the first two being linear estimators. The cross-validated random search is used for hyper-parameter optimization for each of the six classifiers, with k = 5 folds, optimizing for a balanced accuracy score on each test fold to obtain the best parameters on the hyper-search space. Accordingly, these steps are followed for each classifier, for each dataset using only the selected features: split data into k-folds, train k-1 (i.e., four) and test one each round, and record the performance of the classifier for each round on the test subset for each parameter configuration. The classifier with the best performance is selected, together with hyperparameters and its maximum average crossvalidated score, as well as the standard deviation over folds. A detailed step-by-step guide to the full implementation of the algorithm in the first embodiment is provided in Algorithm 1.Algorithm 1 Step-by-step rs-fMRI diagnosis algorithm1 : V rs-fMRI BOLD signal data:2: 1 . V preprocessing strategies:3: (i) with / without global signal correction.4: (ii) with / without band-pass filtering.5: 2. Use V atlas e {AAL, TT}, V preprocessing strategy:6: (i) AAL atlas with 116 brain regions7: (ii) TT atlas with 97 brain regions8: 3. Calculate the two feature representations for each atlas / strategy:9: (i) Static functional connectivity matrix FC10: (ii) Dynamic functional connectivity dFC11 : (I) Use a Gaussian smoothed sliding window over each pair of brain regions to calculate pair-wise dynamic correlations.12: (II) V pair of brain regions, calculate the fraction of no correlation as nwk, and the fraction of strong correlation as nst.13: (III) Use these aggregations for each region pair as the new dynamic functional connectivity feature to create the feature matrix dFC.14: Feature Selection:15: For each feature representation, run a univariate selector to reduce feature space16: For each of the four RFE-CV kernels, find the n features that provide the highest cross-validated balanced accuracy to be used for each of the kernels.17: Classification:18: V classifier, for each configuration of hyper-parameters, for each reduced feature representation:19: (i) Split reduced Xseiect, with n selected features, into k folds, along with labels y.20: (ii) Calculate the cross- validated score for each hyper-parameters’ configuration.21 : (iii) Determine the best hyper-parameters configuration in terms of score for each classifier.22: (iv) Output the best classifier / parameters, along with its used n features.

[0041] To measure the performance of machine learning components used in the first embodiment, different metrics are used to evaluate autism diagnosis, especially when classes are imbalanced on the original dataset or on the output labels. Let TP indicate true positive, TN indicate true negative, FN denote false negative, and FP denote false positive. The following performance metrics are used in this work and defined as follows - Specificity: TN / (FP+TN), Sensitivity (recall): TP / (TP+FN), Accuracy: (TP+TN) / (TP+TN+FP+FN), Balanced accuracy: average of true positive rate (sensitivity) and true negative rate (specificity). SRS assessments from the ABIDE I dataset are used as the ground truth.

[0042] In order to test the significance of the configuration of choice, all model scores were logged for each of the five test stages, labeled as (1 ) feature representation (two) FC or dFC, (2) atlas: (two) TT or AAL, (3) preprocessing strategy: (four) filt global, filt noglobal, nofilt global, or filt noglobal, (4) feature selection kernel of the RFE-CV: (four), and (5) machine learning classifier (six). Thus, the labeled table contains up to 1920 results. Three- factor ANOVA, with the calculation of the sum of squares (SS) for the factors, is used to test the effect of the three major factors pre-machine learning: (1 ) feature representation, (2) atlas, and (3) preprocessing. Full interactions are tested (Table 2), before removing nonsignificant interactions of p > 0.001 and rerunning the test (Table 3). The sum of squares that each factor accounted for did not change greatly due to the nature of the type III sum of squares. In the following part, the effect of each factor is discussed. It is important to note that all configurations (for feature selection, classifier, ... etc.) are considered here, not only high-performing ones, which would be seen as having lower mean accuracy. The effect of each factor is evaluated comparatively, not just quantify its performance, especially as the experiments are paired: less-performing configurations would appear in both groups.Table 2: Multifactorial (3-way) ANOVA results. Feat denotes feature representation, Atls denotes the used atlas, and Strat is one of the four preprocessing strategies. sum_sq, df, F, PR are standard result names of sum of squares, degree of freedom, F_score, and p_value of this F_score, for each factor sum_sq df F PR (>F)C(Feat, Sum) 0.941852 1.0 163.004715 7.364232 x1 0-36C(Strat, Sum) 0.177957 3.0 10.266230 1.053597 x1 0-Q6C(Atls, Sum) 0.149164 1.0 25.815521 4.129224 x1 0-Q7C(Feat, Sum):C(Strat, Sum) 0.117433 3.0 6.774646 1.528091 x10"°4C(Feat, Sum):C(Atls, Sum) 0.013784 1.0 2.385576 1.226282 x1Q-°1C(Strat, Sum):C(Atls, Sum) 0.034962 3.0 2.016924 1.095533 x1Q-°1C(Feat, Sum):C(Strat, Sum):C(Atls, 0.084563 3.0 4.878417 2.210611 xSum) 10"03Table 3: Multifactorial (3-way) ANOVA test results after removing insignificant interactions. sum_sq df F PR (>F)C(Feat, Sum) 0.936316 1.0 160.682470 2.134089 x 10"35C(Filt, Sum) 0.179172 3.0 10.249322 1 .078933 x 10"06C(Atls, Sum) 0.148105 1.0 25.416432 5.062110 x 1Q-°7C(Feat, Sum):C(Filt, 0.117208 3.0 6.704737 1.686921 x 1Q-04

[0043] The choice of the preprocessing strategy (whether to perform global signal correction, and whether to apply band-pass filtering) has a significant effect on the classification performance (p ' 106). From Table 3, with F = 10, we can see that the effect of this factor, although significant, is less important than the other factors. While the difference between the overall mean accuracy for (filt noglobal, nofilt global , nofilt_noglobal) is less than 0.2%, the mean filt global accuracy is higher than 2.5%. Another relevant aspect is that although the max performance is not accompanied by the filt global strategy, the scores are slightly higher across other models of the strategy.

[0044] In the second level of importance comes the used atlas in brain parcellation (TT or AAL), with F = 25.5. The choice of the atlas also has a significant effect on the classification performance (p ' 4 x 107) as we can see in the ANOVA tables. The mean average performance is 2% higher in favor of the AAL atlas. This gives an important indication that the more granular the atlas is (116 areas in comparison to 97 areas), the more informative the features we have (the mean BOLD signal), which leads to better accuracy results.

[0045] As seen in Tables 2 and 3, the feature representation (i.e. , dFC) has the largest effect on the target (accuracy score). This is demonstrated with a high f value of F ' 160 and a highly significant probability of p ' 2 x 1035. On average, the dynamic functional connectivity representation scored an accuracy 5% higher than the conventional functional connectivity.

[0046] With respect to the atlas / feature representation combinations, a trend in performance was found, with (dFC / AAL) being the best performing in general, then (dFC / TT), (FC / AAL), and (FC / TT) having the least scores. In terms of average scores, Isvm kernel of the featureselection is the best performing, and rf is the worst. Both Isvm and Ir classifiers perform the best on average, while rf and Igbm provided the lowest scores. The top achieved balanced accuracy is 98.8% for Isvm feature selection, Isvm classifier, nofilt global preprocessing, AAL atlas, and dFC feature representation. This model configuration (preprocessing, feature representation, atlas, feature selection, classifier, and found best parameters) is elected as the best classifier and is investigated more in the following subsections. The original dFC input is composed of 13,340 features, from which the Isvm-based feature selection elects 840 features, which are used in further processing. The best model is cloned, and five-fold cross-validation using a new random split is made to ensure that the results are not specific to the previous k-fold run. Table 4 shows the results of the best configuration model in terms of the accuracy, sensitivity, specificity, and balanced accuracy ± standard deviations.Table 4: Cross-validated test results of the best configuration modelMetric Accuracy Sensitivity Specificity Balanced Accuracy5-fold value 0.988 ± 0.004 0.987 ± 0.008 0.989 ± 0.007 0.988 ± 0.004

[0047] An interesting finding is that identifying the weak / under-connectivity between brain areas is more important to optimizing model performance than the very strong correlations, with a percentage of wk features = 81.9%, and strong correlations contributing only to the remaining 18.1% of the selected features. The first 10 used features, as well as the top frequent brain regions, are highlighted in Table 5.Table 5: Summary of first 10 selected features as well as the top frequent brain regionsIndex First Selected Features Top Frequent Region FrequencyNames1 Precentral L Rolandic Oper R wk Temporal_Sup_L 242 Precentral L Supp Motor Area L st Postcentral_R 233 Precentral L Fusiform L wk Frontal_Sup_Medial_R 224 Precentral L Parietal Inf L wk Precuneus R 225 Precentral L SupraMarqinal R st SupraMarginal_R 216 Precentral L Precuneus R wk Frontal_Sup_R 217 Precentral L Temporal Sup L wk Cingulum_Ant_R 218 Precentral L Temporal Pole Sup R wk Supp Motor Area L 219 Precentral L Cerebelum Crus2 R wk Angular_R 2110 Precentral R Frontal Inf Tri L wk Hippocampus_R 21

[0048] A comparison of the experimental results of some other existing classification systems are provided in Table 6. The results for the disclosed CAD system are higher than most of the literature for the ABIDE-I dataset. This invention not only establishes better accuracy on a large dataset but also shows comprehensive experimentation of different blocks along the pipeline. Moreover, the ‘average’ performance across the novel dynamicconnectivity feature dFC across all configurations, including other atlas and inferior strategies, is still greater than most of the literature (78.09% average accuracy).Table 6: Comparison of results of certain existing ASD classification methods using the ABIDE-I dataset

[0049] Referring now to the second embodiment, multiple classifiers were again evaluated in determining the most effective classifier that attains the highest accuracy. In the second embodiment, the classifier not only distinguishes between TD and ASD brains but also predicts the severity level of ASD as one of mild, moderate or severe. Forty experiments were conducted using LR, LSVM, LGBM, gradient boosting (GB) and k-nearest neighbor (KNN) machine learning models. The ground truth for these classifications was determined using the Social Responsiveness Scale (SRS) assessment administered by autism experts. As detailed below, LSVM emerged as the classifier achieving the highest accuracy. Bayesian optimization was used to fine-tune the hyperparameters of the LSVM model.

[0050] Tables 7 and 8 present the average and standard deviation of key performance metrics for the LSVM classifier each severity level (mild, moderate, severe) during the training phase and testing phase, respectively. Tables 9 highlights the fifteen most significant functional connectivities between brain regions as indicative of mild, moderate and severe ASD.Table 7: Performance metrics for LSVM with the used features across severity levels on the training phaseTable 8: Performance metrics for LSVM with the used features across severity levels on the testing phaseTable 9: Most significant functional connectivities between brain regions as indicative of mild, moderate and severe ASD

[0051] The disclosed CAD system may be embodied in computer program instructions stored on a non-transitory computer readable storage medium configured to be executed by a computing system. The computing system utilized in conjunction with the CAD system described herein will typically include a processor in communication with a memory, and a network interface. Power, ground, clock, and other signals and circuitry are not discussed, but will be generally understood and easily implemented by those ordinarily skilled in the art. The processor, in some embodiments, is at least one microcontroller or general purpose microprocessor that reads its program from memory. The memory, in some embodiments, includes one or more types such as solid-state memory, magnetic memory, optical memory, or other computer-readable, non-transient storage media. In certain embodiments, the memory includes instructions that, when executed by the processor, cause the computing system to perform a certain action. The computing system also preferably includes a network interface connecting the computing system to a data network for electronic communication of data between the computing system and other devices attached to the network. In certain embodiments, the processor includes one or more processors and the memory includes one or more memories. In some embodiments, computing system is defined by one or more physical computing devices as described above. In other embodiments, the computing system may be defined by a virtual system hosted on one or more physical computing devices as described above.

[0052] Various aspects of different embodiments of the present disclosure are expressed in paragraphs X1 and X2 as follows:

[0053] X1 : One embodiment of the present disclosure includes a method for classifying a neurological disorder comprising receiving neuroimaging data of a subject brain; parcellating the subject brain into a plurality of brain regions; generating feature representations of the subject brain, wherein each feature representation is a measurement of functional connectivity between two of the plurality of brain regions; selecting a subset of the feature representations characteristic of the neurological disorder; and classifying, using a machine learning classifier, the subject brain with respect to the neurological disorder based on the selected feature representations.

[0054] X2: Another embodiment of the present disclosure includes a computer-aided diagnostic system for diagnosis of a neurological disorder, the system comprising at least one non-transitory computer readable storage medium having computer program instructions stored thereon; and at least one processor configured to execute the computer program instructions causing the processor to perform the following operations: receiving neuroimaging data of a subject brain; parcellating the subject brain into a plurality of brain regions; generating feature representations of the subject brain, wherein each feature representation is a measurement of functional connectivity between two of the plurality ofbrain regions; selecting a subset of the feature representations characteristic of the neurological disorder; and classifying, using a machine learning classifier, the subject brain with respect to the neurological disorder based on the selected feature representations.

[0055] Yet other embodiments include features described in any of the previous paragraphs X1 or X2 combined with one or more of the following aspects:

[0056] Wherein the neurological disorder is autism spectrum disorder (ASD).

[0057] Wherein the classifying the subject brain with respect to the neurological disorder comprises classifying the subject brain as one of mild ASD, moderate ASD, severe ASD or typical development.

[0058] Wherein classifying the subject brain with respect to the neurological disorder comprises classifying the presence or absence of the neurological disorder and, if present, classifying the severity of the neurological disorder.

[0059] Wherein the neuroimaging data are functional magnetic resonance imaging data.

[0060] Wherein the neuroimaging data are blood oxygen level dependent (BOLD) signals.

[0061] Wherein the measurement of functional connectivity between two of the plurality of brain regions is the measurement of dynamic functional connectivity between two of the plurality of brain regions.

[0062] Wherein selecting the subset of the feature representations characteristic of the neurological disorder includes selecting the subset of features evidencing one of underconnectivity or over-connectivity as compared to typical development.

[0063] Wherein selecting the subset of the feature representations characteristic of the neurological disorder includes selecting the subset of the feature representations characteristic of the neurological disorder via recursive feature elimination.

[0064] Wherein the machine learning classifier is a linear support vector machine classifier.

[0065] Wherein generating feature representations includes calculating pair-wise correlations between each pair of brain regions in the plurality of brain regions and calculating, for each pair of brain regions, the fraction of strong correlations and weak correlations.

[0066] Wherein calculating pair-wise correlations between each pair of brain regions includes calculating pair-wise correlations between average blood oxygen level dependent (BOLD) signals in each pair of brain regions.

[0067] Wherein feature representations evidencing over-connectivity or under-connectivity as compared to typical development are characteristic of the neurological disorder.

[0068] The foregoing detailed description is given primarily for clearness of understanding and no unnecessary limitations are to be understood therefrom for modifications can be made by those skilled in the art upon reading this disclosure and may be made without departing from the spirit of the invention.

Claims

CLAIMS1) A method for classifying a neurological disorder comprising: receiving neuroimaging data of a subject brain; parcellating the subject brain into a plurality of brain regions; generating feature representations of the subject brain, wherein each feature representation is a measurement of functional connectivity between two of the plurality of brain regions; selecting a subset of the feature representations characteristic of the neurological disorder; and classifying, using a machine learning classifier, the subject brain with respect to the neurological disorder based on the selected feature representations.2) The method of claim 1, wherein the neurological disorder is autism spectrum disorder (ASD).3) The method of claim 2, wherein the classifying the subject brain with respect to the neurological disorder comprises classifying the subject brain as one of mild ASD, moderate ASD, severe ASD or typical development.4) The method of claim 1 , wherein classifying the subject brain with respect to the neurological disorder comprises classifying the presence or absence of the neurological disorder and, if present, classifying the severity of the neurological disorder.5) The method of claim 1, wherein the neuroimaging data are functional magnetic resonance imaging data.6) The method of claim 1, wherein the neuroimaging data are blood oxygen level dependent (BOLD) signals.7) The method of claim 1 , wherein the measurement of functional connectivity between two of the plurality of brain regions is the measurement of dynamic functional connectivity between two of the plurality of brain regions.8) The method of claim 1 , wherein selecting the subset of the feature representations characteristic of the neurological disorder includes selecting the subset of features evidencing one of under-connectivity or over-connectivity as compared to typical development.9) The method of claim 8, wherein selecting the subset of the feature representations characteristic of the neurological disorder includes selecting the subset of the feature representations characteristic of the neurological disorder via recursive feature elimination.10) The method of claim 1, wherein the machine learning classifier is a linear support vector machine classifier.11) A computer-aided diagnostic system for diagnosis of a neurological disorder, the system comprising: at least one non-transitory computer readable storage medium having computer program instructions stored thereon; and at least one processor configured to execute the computer program instructions causing the processor to perform the following operations: receiving neuroimaging data of a subject brain; parcellating the subject brain into a plurality of brain regions; generating feature representations of the subject brain, wherein each feature representation is a measurement of functional connectivity between two of the plurality of brain regions; selecting a subset of the feature representations characteristic of the neurological disorder; and classifying, using a machine learning classifier, the subject brain with respect to the neurological disorder based on the selected feature representations.12) The computer-aided diagnostic system of claim 11, wherein generating feature representations includes calculating pair-wise correlations between each pair of brain regions in the plurality of brain regions and calculating, for each pair of brain regions, the fraction of strong correlations and weak correlations.13) The computer-aided diagnostic system of claim 12, wherein calculating pair- wise correlations between each pair of brain regions includes calculating pair-wise correlations between average blood oxygen level dependent (BOLD) signals in each pair of brain regions.14) The computer-aided diagnostic system of claim 11, wherein feature representations evidencing over-connectivity or under-connectivity as compared to typical development are characteristic of the neurological disorder.15) The computer-aided diagnostic system of claim 11 , wherein classifying the subject brain with respect to the neurological disorder comprises classifying the presence or absence of the neurological disorder and, if present, classifying the severity of the neurological disorder. 16) The computer-aided diagnostic system of claim 11, wherein the neuroimaging data are functional magnetic resonance imaging data.17) The computer-aided diagnostic system of claim 11, wherein the neurological disorder is autism spectrum disorder (ASD); and wherein classifying the subject brain with respect to the neurological disorder includes classifying the subject brain as one of mild ASD, moderate ASD, severe ASD or typical development.

Images